Root Cause Diagnosis of Process Fault Using KPCA and Bayesian

Jan 30, 2017 - Limited diagnostic information from kernel principal component analysis, other online fault detection and diagnostic tools, and process...
0 downloads 12 Views 5MB Size
Article pubs.acs.org/IECR

Root Cause Diagnosis of Process Fault Using KPCA and Bayesian Network H. Gharahbagheri, S. A. Imtiaz,* and F. Khan Faculty of Engineering and Applied Science, Memorial University, St. John’s, Newfoundland and Labrador A1B 3X9, Canada ABSTRACT: This paper develops a methodology to combine diagnostic information from various fault detection and isolation tools to diagnose the true root cause of an abnormal event in industrial processes. Limited diagnostic information from kernel principal component analysis, other online fault detection and diagnostic tools, and process knowledge were combined through Bayesian belief network. The proposed methodology will enable an operator to diagnose the root cause of the abnormality. Further, some challenges on application of Bayesian network on process fault diagnosis such as network connection determination, estimation of conditional probabilities, and cyclic loop handling were addressed. The proposed methodology was applied to Fluid Catalytic Cracking unit and Tennessee Eastman Chemical Process. In both cases, the proposed approach showed a good capability of diagnosing the root cause of abnormal conditions.

1. INTRODUCTION In process industries, there has been a continuing demand for higher quality products and lower product rejection rates, satisfying increasingly stringent safety and environmental regulations.1 Implementation and improvement of an accurate control scheme have been essential over the recent decades in order to meet these ever increasing standards.2 Modern control systems became extremely complex by integrating various functions and components for sophisticated performance requirement.3,4 With such complexities in hardware and software, it is natural that the system may become vulnerable to faults in practice and fault diagnostic tools are required to ensure the process safety and quality of products. The objectives of these tools are early detection of faults and to minimize the impact of a fault on the system.4 In recent years, extensive research has been conducted on process fault detection and diagnosis (FDD). According to the comprehensive review of Venkatasubramanian et al., FDD tools can be divided to model-based methods and process history based methods.5−7 Model based methods require precise mathematical relationship between internal states of the process. Most of the times, it is impossible to have such a precise model.8,9 History based methods uses the data of the process that contain all normal and abnormal condition in the process and implement these data for training and fault detection purpose.10 Although these methods are effective to detect faults early and widely used in process industries, the diagnosis of the faults is not precise. Various residual evaluation methods have been developed to uniquely identify the fault location, for example, generalized likelihood ratio test,11 and structured residuals.12 Zhang et al. (2016) combined Kernel PLS with reconstruction based methods to find the major failure directions.13 However, the inaccuracy in fault diagnosis © 2017 American Chemical Society

still exists and often these methods point toward response variables as the root cause. To overcome the limitations of individual methods and improve the diagnose ability of process faults, hybrid methods have been proposed by researches in recent years. A hybrid framework consists of a collection of methods and utilizes information from several FDDs to overcome the limitations of individual methods. 14−16 This combination of different methods allows one to evaluate different kinds of knowledge in one single framework for better decision making.13 For instance, the analytical methods in model-based fault detection and diagnosis are based on residual generation using parameter estimation; however, the robustness of the model-based methods is often under question since obtaining an accurate model for process, especially for chemical processes, is challenging. To address this problem, Frank proposed the use of analytical methods and integrated them with knowledgebased methods. He combined the analytical methods with expert system approach. Expert system makes use of qualitative models based on available information on the process, facts, and rules. Degree of aging, used tools, and history of operation are examples of expert knowledge. They concluded that knowledge-based methods complement the shortcomings of analytical methods of fault diagnosis.17 An integration of neural network and expert system for fault diagnosis was done by Becraft et al. Once the process fault was localized by neural network, the results were analyzed by a deep knowledge expert system including information on the system structure, functions, and principles of operation.18 Mylaraswamy provided Received: Revised: Accepted: Published: 2054

May 18, 2016 January 26, 2017 January 29, 2017 January 30, 2017 DOI: 10.1021/acs.iecr.6b01916 Ind. Eng. Chem. Res. 2017, 56, 2054−2070

Article

Industrial & Engineering Chemistry Research

Figure 1. Proposed method for root cause diagnosis of process fault.

industrial processes, it is not feasible to monitor all the variables, as such sometimes a fault may originate from an unmonitored variable. To address this issue, Yu et al. investigated the possibility of combining modified ICA and BN for process FDD. The limited diagnostic results of ICA were used as evidence in BN updating, and it was concluded the combined framework of these two methods is a strong tool for FDD purpose for all monitored and nonmonitored variables.21 Despite these researches, there is still more need to research in this area especially because it is somewhat unclear how to use the outcome of data-based FDD methods as an input to the knowledge-based root cause analysis in an automated fashion. Also in these research efforts, the causal relationships in the network were determined using process knowledge and conditional probabilities are assigned based on expert judgment which is subjective as it heavily depends on the knowledge and experience of individuals. Furthermore, there is ambiguity in whether one should use normal or faulty data for network training and details on preprocessing of training data is not discussed.19,21 Some application challenges arise due to the inherent limitations of Bayesian network. A BN in its original form is an acyclic directed graph. In chemical processes, however, cyclic loops appear due to material and heat integration, recycle streams, as well as information flow paths due to feedback control. As such, BNs needs to be adopted to represent chemical processes adequately.

a brief comparison of the various diagnostic methods to highlight the inadequacy of individual methods and underscored the need for collective problem solving. They proposed a Dkit based hybrid method of neural network for detection of a fault and a signed directed graph for diagnostic action.16 Zhao et al. proposed a wavelet-sigmoid basis neural network for dynamic diagnosis of failure in hydrocracking process.15 Process measurements are very noisy, and there is uncertainty in the relationship between process variables. A BN model is an excellent tool to characterize processes with stochastic uncertainty using conditional probability-based state transitions.19 Hence, it can be adopted to identify the propagating probabilities among different measurement variables so as to determine the operating status of processes and diagnose the root causes of abnormal events.19 Researchers have used BN for improving process fault diagnosis in different ways. Because of the stochastic nature of process variation, a false alarm may be generated in the monitoring system while the process is operating in normal condition. To address this problem, in the methodology proposed by Dey et al., data from multiple sensors were combined through a causal belief network to estimate probabilistic diagnosis of root cause of the process fault. They showed that the posterior probability of each node that shows the status of the node can be updated from evidence using Pearl massage passing algorithm.20 Yu et al. proposed a Bayesian inference-based abnormality likelihood index to detect a process fault. In diagnosis phase they utilized dynamic Bayesian probability and contribution indices.19 In 2055

DOI: 10.1021/acs.iecr.6b01916 Ind. Eng. Chem. Res. 2017, 56, 2054−2070

Article

Industrial & Engineering Chemistry Research Our research aims to fill these knowledge gaps with view to developing comprehensive methods that can precisely diagnose the root cause of process fault and help the operator to take corrective actions. We propose a new methodology through integration of diagnostic information from various single variable and multivariable diagnostic tools using BN. We also address the application difficulties of BN related to process fault diagnosis. We used Granger causality and transfer entropy to determine the causal relationships between process variables. A detailed methodology for estimating conditional probabilities between variables is also proposed in this paper. Cyclic networks were dealt with through transformation of cyclic BN to acyclic-BN using pseudonodes. We demonstrated the efficacy of the proposed methodology through two case studies, i.e., FCC and Tennessee Eastman.

2. METHODOLOGY: BN-BASED PROCESS MONITORING APPROACH The overall picture of the methodology is given in Figure 1. The main components of the proposed methodology are kernel principal component analysis (KPCA) model, sensor fault detection module, and BN. In an online setting, KPCA will act as a primary tool for detecting process fault. The reason for using KPCA for online fault detection is KPCA can model process nonlinearities well. Also, from an application point of view, compared to other nonlinear methods less trial and error is required for building KPCA model because simple Kernel functions can fit a wide range of nonlinearities. Having detected the fault, the next step would be to diagnose the root cause of the fault and the propagation pathway. The fault may be originated among internal states, or it may be because of a sensor malfunction. It is important to isolate sensor faults using sensor validation module because these kind of faults breakdown the correlation between variables and the causality networks do not work effectively. A sensor check module can be designed using simple rule-based algorithms or more sophisticated algorithms such as bank of Kalman filter with weighted sum of squared residual (WSSR).22 To keep the methodology simple, we used a rule-based algorithm. If the sensors are working properly, the failure should be among process internal states. The BN is more appropriate to detect disturbance type faults. Thus, if it is a sensor fault the algorithm stops at this point. Otherwise, it will proceed to determining the root cause of the fault. The average contributions calculated from the KPCA is used to diagnose preliminarily the causes of faults. This information is passed on to the trained BN as evidence. The trained BN based on its causal relations and conditional probabilities determines the true root cause of the fault. The training of the KPCA and BN is done in an off-line mode. Important steps in building a KPCA model are data normalization, determining the number of PCs, residual generation using training data set. BN has two components: construction of causal network and estimating the conditional probabilities. Granger causality and transfer entropy was applied for estimation of mutual causal relationship between variables and for construction of network. Knowledge of the process was used to verify the constructed network. If the network contained a loop, a dummy duplicate node was created for one of the variables involved in the loop because Bayesian network cannot update cycles. Conditional probabilities among different nodes are the quantitative part of the network which was calculated using maximum likelihood method. Our objective was to reflect the causal relationships

Figure 2. Commonly observed sensor faults: (a) FLAT LINE fault, (b) SPIKE fault, (c) NOISE fault.

among the variables due to abnormal events as such we used the residuals from the KPCA to calculate the conditional probabilities. The residuals contain only the abnormal process variations thus conditional probabilities calculated from the residuals better reflect the propagation pathways of the faults. More detail on each section is presented in following sections: 2.1. Kernel Principal Component Analysis (KPCA). KPCA is an extension of PCA to deal with nonlinear data set. In KPCA, nonlinear data can be converted to linear form through high-dimensional mapping. For example, a data set x and z, which are not separable in current space, are linearly separable in nonlinear hyperplane with features Φi(x) and Φj(z). Thus, KPCA is a two-step method: calculation of covariance matrix and dot products of variables in feature space, and singular value decomposition of covariance matrix in feature space. Finding the exact feature space is not straightforward, and calculation of the dot products in the feature space can be prohibitive due to calculation complexities. Instead of explicitly transforming variables to feature space, dot 2056

DOI: 10.1021/acs.iecr.6b01916 Ind. Eng. Chem. Res. 2017, 56, 2054−2070

Article

Industrial & Engineering Chemistry Research

Table 2. Mutual Transfer Entropy Values between Nodes on Dissolution Tank System Water in Solid Level

Water in

Solid

Level

Concentration

NA NA NA

NA NA NA

0.16 0.14 NA

0.25 0.19 0.02

Figure 3. Process flow diagram of dissolution tank system. Figure 4. Constructed network for dissolution tank system using Granger causality and transfer entropy.

product vectors in feature space is calculated using Kernel function. The dot product of the transformed variables is given by the following equation:

k(x , z) = Φi(x) ·Φj(z)

should be minimum; however it is not zero because there is always some process normal variation and noise in a process. The residuals from the KPCA module were used to update the probability of the corresponding nodes in the BN. The residual components are calculated from the following:

(1)

The function k(·, ·) is called the kernel function, and there exist several types of these functions. Some popular kernel functions are polynomial functions, Gaussian functions, and sigmoidal functions.23,24 The fault detection in KPCA is done using Hotelling T2, which is similar to PCA. The contributions of each variable reflect some useful information in diagnosis. However, unlike PCA contribution of KPCA model cannot be calculated easily because of the nonlinear transformation in KPCA. Alcala and Qin proposed reconstruction-based contribution (RBC) to overcome the aforementioned shortcoming and to estimate contribution of each variable.24,25 The procedure to estimate fault free data by eliminating the effect of fault from faulty data is defined as reconstruction.

zi = xi − ξifi

t ̃ = Pf T Φ

(3)

T

where Pf is related eigenvectors corresponding to remained PCs, which are related to residuals and Φ is the mapped vector of observation space. The calculation of residual from eq A3 is not possible because of high dimensionality in feature space. Here we calculated residuals from the difference between faulty data and fault free data, which is obtained from reconstruction: ei = xi − zi (4) More details and the mathematics behind the theory are given in the Appendix. 2.2. Sensor Failure Module. In the proposed method, a rule-based sensor check module has been designed after KPCA unit to isolate sensor malfunction. According to Sharma et al. the following failures are probable in a sensor:27 • FLAT LINE: The sensor reports a constant value for a large number of successive samples (Figure 2a). • SPIKE: A sharp change in the measured value between two successive data points (Figure 2b).

(2)

where zi is fault free data, xi is faulty data, ξi is the direction of fault, and f i is the magnitude of fault. RBC considers the reconstruction of a fault detection index (T2 or SPE) along the direction of a variable as the variable’s contribution for that fault. In other words, the objective of RBC is to find f i of a vector with direction ξi such that the fault detection index of the reconstructed measurement is minimized. The reason for minimization is that for fault free data the detection index

Table 1. Result of Granger Causality Analysis for Dissolution Tank System Water flow Water flow RPM Level Concentration

0.7632 0.6466 53.332 0.7524 13.644 0.8323

RPM

Level

Concentration

66.65 0.1253

53239.33 0.0000 3424.43 0.0000

4554.66 0.0000 76434.5 0.0000 949.43 0.5322

0.7613 0.0934 6347.3 0.0784

361.54 0.1125 2057

Chi-sq Prob. Chi-sq Prob. Chi-sq Prob. Chi-sq Prob. DOI: 10.1021/acs.iecr.6b01916 Ind. Eng. Chem. Res. 2017, 56, 2054−2070

Article

Industrial & Engineering Chemistry Research

2.3. Bayesian Network Construction. A BN is utilized to determine the fault origin and the pathway in which the fault is propagated. BN has two components: the causality network and conditional probabilities. Typically, in most cases the causal relationship and conditional probabilities are assigned based on process knowledge. In this study, we show how to use process data to complement process knowledge. To construct the network, the mutual cause and effect relationship between variables should be determined. We employed Granger causality and transfer entropy to extract causal relationships among variables from process data. The detailed explanation on causality analysis and construction of the BN is described below: 2.3.1. Causality Analysis Based on Granger Causality. Wiener introduced the concept of history based causality, and later in 1969 Granger formulated it to show the cause and effect relationship among different variables in any system.27−29 According to Granger, a variable y Granger causes the other variable x if incorporating the past values of x and y helps to better predict the future of x than incorporating the past values of x alone.27 Considering two time series x and y, there are two different linear regression model. One is a restricted model in which the prediction of x at time k is possible using the information on the past values of x: p

xk =

∑ μixk− i + εxk

(5)

i=1

where xk is the value of x at time k; xk‑i is the i-lagged value of x; μ is the regression coefficient; p is the number of time lagged variables considered; and ε denotes the residual series for constructing xk. The second model is unrestricted model in which the prediction of x at time k is possible using the past information on both x and y as follows:

Figure 5. Process value and generated residuals of dissolution tank system: (a) Level, (b) water flow rate.

p

xk =

q

∑ γixk− i +

∑ βi yk− j + ηxk

i=1

j=1

(6)

where xk is the x time series at time k; xk‑i and yk‑j are respectively the i-lagged x time series and j-lagged y time series; γ and β are the regression coefficients; p and q are the amount of lag considered or model order; and η denotes the unrestricted model residual at time k. The μ, γ, and β parameters are calculated using least-squares method. A small value of p or q leads to poor model estimation whereas large values result in problem of overfitting. Akaike Information criterion (AIC)30 and Bayesian Information criterion (BIC)31 are two criteria that are used to determine the model order. AIC and BIC are given as follows:

Figure 6. Loop handling in Bayesian network.

• NOISE: The variance of the sensor readings increases. Unlike SPIKE faults that affect a single sample at a time, NOISE faults affect a number of successive samples (Figure 2c). To keep things simple, we used the following rules to detect these faults: • When the difference between two consecutive observations is less than a very small value and this happens for five consecutive observations, the sensor is suspected to FLAT LINE failure. • When the difference between two consecutive observations is a very large number in comparison with the standard deviation of data, the sensor is suspected to SPIKE failure. • If in any data window of N data points, q measured values of a sensor exceed the threshold, the sensor is suspected for NOISE failure.

AIC(p) = ln(det(Σ)) +

2pn2 T

(7)

BIC(p) = ln(det(Σ)) +

ln(T )pn2 T

(8)

where Σ represents the noise covariance matrix, n is the number of variables, and T is the total number of observations. The value of p, which maximizes the value of AIC or BIC, is usually selected as model order. When the variability of the residual of the unrestricted model is significantly reduced with that of a restricted model, then there is an improvement in the 2058

DOI: 10.1021/acs.iecr.6b01916 Ind. Eng. Chem. Res. 2017, 56, 2054−2070

Article

Industrial & Engineering Chemistry Research

of-thumb” approach. For a q-dimensional multivariate case with independent probability distribution for each variable, the estimated PDF is given by34

prediction of x due to y. In other words, y is said to Granger cause x. This improvement can be measured by the F statistic: F=

(RSSr − RSSur )/q ∼ F(p , T − p − q − 1) RSSur /(T − p − q − 1)

P(x1 , x 2 , ..., xq) =

(9)

where RSSr is the sum of the squares of the restricted model residual, RSSur is the sum of the squares of unrestricted model residual, and T is the total number of observations used to estimate the model. F statistics approximately follows F distribution with degrees of freedom p and (T − p − q −1). If the F statistics from y to x is significant, then the unrestricted model yields a better explanation of x than does the restricted model, and y is said to Granger cause x.31 The threshold of F statistics was set at 5% significance level. 2.3.2. Causality Analysis Based on Transfer Entropy. For two variables with sampling interval of τ, xi = [xi, xi‑τ, ... ; xi‑(k‑1) τ] and yi = [yi, yi‑τ, ..., yi‑(l‑1) τ], information transferred from y to x is defined as follows:32 t (x | y ) =



P(xi + h , xi , yi ) ·log

xi + h , xi , yi

(10)

n

1 −1/2v 2 e 2π

P(x) =

1 N×d

N

⎛ x − xi ⎞ ⎟ d ⎠

i=1

∑ j = 1 [xj(i) = x] n

=

count (x) n

(15)

n

(i) (i) P(x , y) ∑ j = 1 [xj = x and yj = y] P(x|y) = n P(y) ∑ [y (i) = y] j=1

(11)

=

j

count j (x and y) count (y)

(16)

This is a very natural estimate and equal to the number of times state x and y both are seen within the threshold; count the number of times the label y is seen within the threshold; then take the ratio of these two terms.37 In a process system, propagation path for faulty and normal variation is often not the same.38 The conditional probability in essence should reflect the causal relations between variables under faulty conditions. In order to keep only the variation of abnormality and noise and to mitigate the effect of process variation on the conditional probability calculation, the residuals from KPCA analysis were used for calculation of conditional probabilities (please see the illustrative example and Figure 5). Also, the residuals follow Gaussian distribution more closely compared to the raw data, as such gives better maximum likelihood estimates. In other words, conversion of data to residuals will mitigate process variations in variable values while keeping the causality of fault information inside, consequently conditional probabilities estimated from residuals reflect the causal relations for faulty variables more accurately. 2.3.4. Construction of BN: An Illustrative Example. We will illustrate the network construction using Granger causality and transfer entropy and also estimation of conditional probabilities, for a simple dissolution tank system.39 In this system, a pure solid crystal is dissolved in a tank with water (Figure 3).

(12)

∑ K ⎜⎝

(14)

[x(i) = x] is 1 if x(i) = x; otherwise, it is equal to zero. Hence, Σni=1[x(i) = x] = count (x) is simply the number of times that the state x is seen in the training set, or number of times that state x is inside threshold that have been previously determined. Similarly, the MLE for the P(x|y), x and y∈{1...k}, takes the following form:

Therefore, a univariate PDF can be estimated by p( x ) =

i=1

where ds = for s = 1, ...,q. Process variables are highly correlated, as such eq A14 may introduce errors in the PDF estimates. 2.3.3. Estimation of Conditional Probabilities. Besides the causal network, a BN contains conditional probability tables for all nodes. These values quantify the amount of influence each node receives from its parents. In this paper, calculation of conditional probabilities was done by maximum likelihood estimation (MLE). Suppose a sample consisting of m variables and n observations. We write xj(i) for the j’th observation of i-th variable. Given these definitions, the MLE for P(x) for x∈{1... k} takes the following form (k is the number of states):

t(y → x) is causality measure and is derived by comparing the influence of y to x with influence of x to y. If t(y → x) is negative, then information is transferred from x to y. Because at first there is no knowledge about which node is cause and which one if effect, choosing these nodes inversely will results in negative value. The advantage of using transfer entropy is that it is a model free method and can be applied to nonlinear data. It has already been proved to be very effective in capturing process topology and process connectivity.36 But it suffers from a large computational burden due to the calculation of the PDFs.33 Histograms or nonparametric methods, e.g., kernel method, can be used to estimate the PDF. The Gaussian kernel function is used to estimate the PDF, which is defined as follows:34 K (v ) =

⎛ x1 − xi1 ⎞ ⎟... ⎝ d1 ⎠

cσ(xi,s)i N= 1N−1/(4+q)

where P(.) denotes the probability density function (PDF) and h is the prediction horizon. k and l show the length of time series. Transfer entropy represents the measure of information transfer from y to x by measuring the reduction of uncertainty while assuming predictability.34 It is defined as the difference between the information about a future observation of x obtained from the simultaneous observation of past values of both x and y, and the information about the future of x using only past values of x. It was shown that the parameter values can be chosen as τ = h ≤ 4, k = 0, and l = 1 for the initial trial.35 Using the above definitions, direction and amount of net information transfer from y to x is as follows: t(y → x) = t(x|y) − t(y|x)

N

∑ K⎜

⎛ xq − xiq ⎞ ⎟⎟ K ⎜⎜ ⎝ dq ⎠

P(xi + h|xi , yi ) P(xi + h|xi)

1 N × d1...dq

(13)

where N is the number of samples, x is an independently drawn sample, d is the bandwidth chosen to minimize the error of estimated PDF. d is calculated by d = cσN0.2 where σ is variance and c = (4/3)0.2 ≈ 1.06 according to the “normal reference rule2059

DOI: 10.1021/acs.iecr.6b01916 Ind. Eng. Chem. Res. 2017, 56, 2054−2070

Article

Industrial & Engineering Chemistry Research

Figure 7. Schematic illustration of FCC process.

The flow of water to the tank is under control. Also, the solid crystals are fed from a hopper to the tank through a rotary feeder. The control objective of this process is to maintain the level of water and concentration of crystal in the tank to desired set point. However, these two parameters are subjected to abnormal changes due to disturbance in the solid discharge. There are four variables in this process: water flow to the tank, RPM of the rotary feeder, level of water, and concentration of solid crystal in the tank. These four variables are taken as nodes in a network, and their values within the operation of the process will be analyzed by Granger causality and transfer entropy for network construction. We used a statistical software Eviews to perform Granger causality analysis. The results of Granger causality analysis are shown in Table 1. In this table based on the F statistics when the prob < 0.05, it rejects the null hypothesis and indicates the variable in the row has influence on the variable in column. For example, based on the first row of this table, water flow has influence on level and solid crystal concentration. Also, the second row shows that the RPM has influence on level and solid crystal concentration as well. The network constructed based on this method is shown Figure 4. Also we did causality analysis of the same system using transfer entropy. The results of transfer entropy are given in Table 2. In this model, there is time lag between inputs and outputs. In calculation of transfer entropy, several time delay values in the range of [−5:5] were considered and the value with the maximum transfer entropy was selected. In Table 2, the value of transfer entropy indicates that the variables in rows have influence on the variables in columns. However, by comparison of the network constructed by Granger causality and transfer entropy, there is a discrepancy in the results. Based on transfer entropy, the level of the tank has ab effect on solid crystal concentration; however, Granger causality does not confirm such a relation. The results of Granger causality are more accurate than that of transfer entropy because in calculation of conditional probability values of transfer entropy test, histograms were used that introduced some error to the calculation. To rectify the results, a threshold of 0.1 was set for transfer entropy values; only values above the threshold indicate that there is a significant causal relationship between the variables. The threshold value is subjective and varies from process to process and also the noise levels in the signals. In

this case we used our knowledge of the process to eliminate weak causalities observed in the data based analysis. Next, we calculated the conditional probabilities for the system. Figure 5 shows 2000 observations for level and water flow rate and their corresponding residuals for the dissolution tank system. The training data contains two kinds of variation in the system: normal process variation and variation due to abnormality and noise. As can be seen in this figure, in the case of raw data there are many fluctuations in data and sometimes they exceed the threshold but in reality these were normal operational changes in the process. However, this is not the issue in residuals as the residuals do not contain process variations. In the case of level, 644 samples exceeded the threshold but in the corresponding residuals only 91 samples exceeded the threshold. Also, in the case of water flow rate, 782 samples exceeded the threshold but in the corresponding residuals only 114 samples exceeded the threshold which is reasonable because the selected PCs for this system show 85% of variation in the system. Considering one standard deviation around the mean as normal threshold, the probability of level of the tank being in normal state is 68% for original data and is 95% for residuals. Also, the probability of water flow rate being in normal state is 61% for original data and is 94% for residuals. This is because in original data some process normal variations are incorrectly considered as fault. So, conditional probability values calculated from the residuals are more accurate than that of from raw data. 2.4. Loop Handling. Because BN is acyclic network, while applying it for chemical processes with feedback controller or recycles, a special treatment is required to convert the network form cyclic to acyclic form. To capture the feedback effect in an acyclic network, we designed a duplicate dummy point as the feedback effect in recycle or controller. For example, in Figure 6a, there is a causal relationship from Xi to Xo. Also, based on the recycle loop, there is a causal relationship from Xo to Xi as well. This loop has been treated as Figure 6b and a dummy variable has been dedicated to variable Xi. It is obvious in Figure 6b variable Xi has effect to variable Xo in the continuous line and also variable Xo has effect to variable Xi in the dash line. 2.5. Fault Diagnosis Using BN. The origin of Bayesian statistics is Bayes’ theorem, which is an equation relating conditional and marginal probabilities: 2060

DOI: 10.1021/acs.iecr.6b01916 Ind. Eng. Chem. Res. 2017, 56, 2054−2070

Article

Industrial & Engineering Chemistry Research P(S|E) =

p(E | S )p(S ) p(E )

(17)

where S and E are event in a sample space and p(.) is probability distribution. Posterior probability of each node is updated from evidence. BN-based root cause diagnosis is an inference process where the evidence is received at the child node. The evidence in this case is the contributions of the KPCA. The variable which has the maximum contribution is suspected as the cause and its evidence is set to 100%. When new evidence is introduced into the network, each node updates its own belief, based on message received from its children. The Bayesian rule for updating will be expressed as P(S|E) = p(E|S)p(S)

(18)

This equation expresses that the posterior belief is proportional to the prior belief times the likelihood.

3. APPLICATION OF PROPOSED METHODOLOGY To demonstrate the effectiveness of the proposed hybrid technique based on KPCA and BN, the methodology was Table 3. Measured Variable of FCC Process No.

Symbol

Variable

1 2 3 4 5 6 7 8 9 10 11 12 13 14

Tair Psi T1 P4 DP Fair P6 T3 T2 Tr Treg Tcyc Csc Crgc

Air temperature Coking factor Fresh feed entering furnace Reactor pressure Differential pressure Air flow rate to regenerator Regenerator pressure Furnace temperature Fresh feed entering to riser Riser temperature Regenerator temperature Cyclone temperature Coke frac. In spent catalyst Coke frac. In regenerated catalyst

Table 4. Fault Scenarios in FCC Scenario no.

Fault description

1 2

5 °C in atmosphere temperature Gradual increase of 10 °C in fresh feed temperature

Figure 8. (a) Hotelling T2 plot for disturbance in ambient air temperature in FCC. (b) Hotelling T2 plot for disturbance in feed temperature in FCC. (c) Contribution plot for disturbance in ambient air temperature in FCC. (d) Contribution plot for disturbance in feed temperature in FCC.

applied to two case studies, one without recycle (FCC) and one with recycle (Tennessee Eastman): 3.1. Fluid Catalytic Cracking. FCC process converts a number of heavy hydrocarbons with different molecular weights to more lighter and more valuable hydrocarbons. The heavy hydrocarbons come from different parts of refinery and are diverse in chemical properties. A schematic illustration of the FCC unit reactor/regenerator section is shown in Figure 7. The monitored variables are listed in Table 3. There are three inputs to the system: fresh feed temperature, feed coke factor, and atmosphere temperature. Feed coke factor was maintained at 1.05 during the entire simulation. Two disturbances were introduced to the system through the other two inputs as described in Table 4.40 Fault Scenarios. Step Disturbance in Ambient Temperature. The first faulty scenario begins with normal operation

for 1000 s and then is followed by a 5 °C increase in ambient air temperature for the remaining 4000 s. The sampling time of the data generation is 1 s. The first 1000 fault free samples were used as training data. Gaussian function was selected as kernel function for KPCA. All data are normalized around zero. Five principal components were selected that explain 85% of the variations in the system. The value of threshold was calculated as 9.71 at 95% of confidence level from the following equation: TThreshold 2 =

k(n − 1) F (k , n − k , α ) (n − k )

Where the F(k, n − k, α) corresponds to the probability point on the F-distribution with (k, n − k) degrees of freedom and 2061

DOI: 10.1021/acs.iecr.6b01916 Ind. Eng. Chem. Res. 2017, 56, 2054−2070

Article

Industrial & Engineering Chemistry Research

is detected, next step is to diagnose its root cause. The plot in Figure 8c depicts the contribution plots for this scenario. As it is obvious in Figure 8c, when there is a disturbance in ambient temperature, the contribution plot cannot exactly diagnose the root cause of the fault, rather point toward few variables involved in the fault. However, regenerator pressure has the most contribution in this abnormal event, i.e., this variable has the highest variation among the variables in the propagation pathway of the fault. Having detected the fault, the first task is to determine whether this is a sensor fault or not. After all probable sensor faults were tested, it was found that all sensors are working properly. Next, we use BN to diagnose the root cause. To construct the network, all monitored variables and disturbances were considered as nodes. The cause and effect relationship between variables is determined by both causality analysis and process knowledge. Granger causality and transfer entropy were used to investigate the cause and effect relationship between variables and process knowledge that includes process flow diagram and expert knowledge were used as a confirmation to the constructed network. The conditional probabilities were obtained from historical data using MLE. In the historical data, both process faults were simulated (Table 4) and the data in training set are samples of all possible abnormal events. As explained in Section 2.3.3 and 2.3.4, the normal variation of process introduces inaccuracy in calculation of conditional

Table 5. (a) Granger Causality and (b) Transfer Entropy Analysis for Selected Variables in FCC Process (a) T1 T1 T3

4.8931 0.0866 3.9197 0.1409

T2

T2

3536.90 0.0081

62441.16 0.0000 399.07 0.0000

2.1850 0.3354 (b) T1

T1 T3

T3

Chi-sq Prob. Chi-sq Prob. Chi-sq Prob.

T3

T2

0.24

0.38 0.42

significance level α. k is the number of principle components and n is the number of observations. The Hotelling’s T2 and contribution plots of the KPCA analysis are shown in the Figure 8. Based on Figure 8a, the Hotelling’s T2 identified departure from process normal condition. This plot depicts a successful detection of the abnormal condition. As can be seen, there is a long delay associated with the detection of this disturbance. It is because the magnitude of the variation in ambient temperature is not big enough to affect the process in a short time. Once the fault

Figure 9. (a) BN for FCC process and (b) fault propagation pathway in FCC process for fault in ambient air temperature. 2062

DOI: 10.1021/acs.iecr.6b01916 Ind. Eng. Chem. Res. 2017, 56, 2054−2070

Article

Industrial & Engineering Chemistry Research

Figure 10. (a) BN for FCC process and (b) fault propagation pathway in FCC process for fault in fresh feed temperature.

temperature (T3). The corresponding network for these three variables is depicted in Figure 9a. When there is a variation in ambient temperature, based on the contribution plot (Figure 8c), the regenerator pressure (P6) has the highest variation among all monitored variables in the process. We take this variable as evidence [Pevidence node7 (state 0) = 100%] for BN updating and for further analysis in BN to find out the propagation path and the true root cause of the fault. The updating of the network was done using GeNIe software. The updated network for the faulty condition is shown in Figure 9a. The state 0 shows the faulty state for each variable and state 1 shows the normal state. Based on this figure, ambient air temperature has a probability of 88% to be in faulty state [PTair(state = 0)= 88%], pointing this node as the most potential root cause of the fault and this variation propagated in this system through regenerator temperature and will be reflected in regenerator pressure. The propagation path is given in Figure 9b. Ramp Disturbance on Feed Temperature. The second faulty scenario begins with normal operation for 1000 s and then is followed by a 10 °C ramp in fresh feed temperature for the remaining 4000 s. The KPCA model construction is same as the first abnormal event. Figure 8b shows the departure of T2 values beyond the threshold that shows the successful detection of fault. After testing all sensors and being confident about their function, the root cause of the fault should be diagnosed among the internal states of the process. In the corresponding

probabilities. Therefore, the residuals of KPCA were used for conditional probability estimation. The network construction was performed using Granger causality and transfer entropy. The results are shown in Table 5. To avoid complexity, we do not bring the results for all variables and we focus on specific part of the process, for example the furnace area, and will construct the network for three variables in this region; i.e., fresh feed temperature (T1), preheated feed temperature (T2), and furnace temperature (T3). In Table 5a where the probability is less than 0.05, the variable in the row has effect on the variable in the column. The first raw of this table shows that fresh feed temperature has effect on furnace temperature and preheated feed temperature which in entering to riser. Also, furnace temperature has influence on preheated feed temperature. The mutual transfer entropy values for this abnormal condition are given in Table 5b. These results are in accordance with the results of Granger causality, which all verify the causal effect elicited from process knowledge. Based on process knowledge, heat balance around furnace in steady state shows that the enthalpy of the preheated fresh feed temperature is equal to the enthalpy of the fresh feed entering to the furnace plus the net amount of enthalpy given to the fresh feed by the furnace. At constant pressure in the furnace, the enthalpy is a function of temperature and consequently the preheated feed temperature (T2) is affected by fresh feed entering to the furnace (T1) and furnace 2063

DOI: 10.1021/acs.iecr.6b01916 Ind. Eng. Chem. Res. 2017, 56, 2054−2070

Article

Industrial & Engineering Chemistry Research

Figure 11. Schematic diagram of Tennessee Eastman process.

contribution plot (Figure 8d), the furnace temperature has the highest contribution to T2 that means after the fault initiated this variable has the highest value of variation due to abnormality in the process. Based on this methodology, this variable will be used as evidence node in the BN [Pevidence node8 (state 0) = 100%]. Based on Figure 10a, the updated network shows that the fresh feed temperature is the root cause of the abnormality in the process [PT1(state = 0) = 62%]. The propagation path is not lengthy and contains just two nodes (Figure 10b). 3.2. Tennessee Eastman Chemical Process. To illustrate further the applicability of the proposed method, the methodology is applied to benchmark Tennessee Eastman chemical process. The process consists of five major units: a reactor, condenser, compressor, separator, and stripper; and, it contains eight streams: A, B, C, D, E, F, G, and H. The flow diagram of this process is shown in Figure 11. It consists 41 measured variables and 12 manipulated variables. Among measured variables, 22 variables are continuous process variables and 19 variables are related to composition measurements. The 22 continuous process variables are shown in Table 6 that are the main focus of this research. There are 20 potential faults in this process, among them we concentrate our study on those that are mentioned in Table 7.5 Fault Scenarios. Sudden Loss of Feed A. The first faulty scenario is a loss in feed A at 500 s. This variation will affect the concentration of all components in the reactor. Consequently, this may change the process parameters in reactor and downstream units. A training set consisting of 500 samples from the normal data was used to develop the KPCA model. Gaussian kernel function was selected for linearization of data in KPCA. Four principal components were selected that

Table 6. Measured Variables in Tennessee Eastman Variable

Description

XMEAS(1) XMEAS(2) XMEAS(3) XMEAS(4) XMEAS(5) XMEAS(6) XMEAS(7) XMEAS(8) XMEAS(9) XMEAS(10) XMEAS(11) XMEAS(12) XMEAS(13) XMEAS(14) XMEAS(15) XMEAS(16) XMEAS(17) XMEAS(18) XMEAS(19) XMEAS(20) XMEAS(21) XMEAS(22)

A Feed (stream 1) D Feed (Stream2) E Feed (Stream 3) Total Feed (Stream 4) Recycle Flow (Stream 8) Reactor Feed Rate (Stream6) Reactor Pressure Reactor Level Reactor Temperature Purge Rate (Stream 9) Separator temperature Separator level Separator pressure Separator Underflow (Stream 10) Stripper Level Stripper Pressure Stripper Underflow (Stream 11) Stripper Temperature Stripper Steam Flow Compressor Work Reactor Cooling Water Outlet temperature Separator Cooling Water Outlet Temperature

Table 7. Fault Scenarios in Tennessee Eastman Chemical Process Fault no.

Fault description

IDV(6) IDV(12)

Sudden loss of flow in feed A Random variation in condenser cooling water inlet temperature

2064

DOI: 10.1021/acs.iecr.6b01916 Ind. Eng. Chem. Res. 2017, 56, 2054−2070

Article

Industrial & Engineering Chemistry Research

and these values exceed the threshold instantaneously when the abnormality initiated indicating that KPCA is able to detect the feed A loss quickly. Next, we confirmed using the sensor check module that all sensors are functioning well and moved on to detect the propagation pathway of fault in internal states using BN. The network construction and parameter estimation for Tennessee Eastman process was similar to FCC process using Granger causality and transfer entropy and was verified by process knowledge. Unlike FCC process, Tennessee Eastman contains a loop that makes the updating of the network difficult. To deal with this issue, we designed a duplicate dummy point (pseudo point) for one of the variables involved in the loop. The constructed network in Figure 13a shows that recycle flow, XMEAS(5), is involved in a loop in the network and was duplicated. One of these nodes is functioning as a parent node and the parameter estimation for this node was conducted like a parent node. The other node is like a child node and conditional probability values were considered for this node. Based on process knowledge, it is obvious that variation in the feed A will affect the reactor feed rate because the inlet to the reactor in summation of A, D, and E feeds. Because these feeds are gaseous, any variation in feed rate will affect the pressure in the reactor. Also, it will affect the conversion due to change in residence time in the reactor which affects the temperature of the reactor due to the exothermic nature of the reaction. The contribution plot of each variable for the fault scenario is given in Figure 12c. XMEAS(1), XMEAS(7), XMEAS(9), and XMEAS(21) have a high contribution in this faulty event; however, XMEAS(21) which is the reactor cooling water outlet temperature has the highest contribution and will be used as evidence in Bayesian networkbased fault diagnosing module [PXMEAS(21) (state 0) = 100%]. The result of the updated network is shown in Figure 13a. As can be seen in this figure, a variation in the feed A propagates though the network, affecting all nodes in propagation pathway, and eventually show up on XMEAS(21), as the last node in the propagation pathway. The probability of feed A to be in faulty state is 78%. In this abnormal scenario, between the true root cause, XMEAS(1), and the faulty monitored variable, XMEAS(21), there are three intermediate variables (Figure 13b). Fault in Condenser Cooling Water Inlet Temperature. The fault initiated with a random variation in condenser cooling water inlet temperature at 500 s. This variable is not among measured variables. Because the training data for both faulty scenarios in Tennessee Eastman are the same, the KPCA model for the second faulty scenario is the same as the first one. Figure 12b shows the departure of T2 values beyond the threshold (threshold = 21.5). The random variation fault in the process variable was detected with a delay because the condenser cooling water inlet temperature is further downstream in the process and has less impact on the system compared to inlet feed. In this abnormal event, XMEAS(9), XMEAS(11), XMEAS(14), and XMEAS(21) have the most contribution in Hotelling T2 (Figure 12d); however, XMEAS(14), which is separator underflow, has the highest contribution and this node was considered as evidence node in the BN [PXMEAS(14) (state 0) = 100%]. The updated network shows that separator temperature, XMEAS(11), is the root cause of this abnormality, having a value of 0.78 as the probability for faulty state (Figure 14a). In this abnormal event, a variation on the condenser

Figure 12. (a) Hotelling’s T2 for the fault IDV(6) in Tennessee Eastman process. (b) Hotelling’s T2 for the fault IDV(12) in Tennessee Eastman process. (c) Contribution plot for the fault IDV(6) in Tennessee Eastman process. (d) Contribution plot for the fault IDV(12) in Tennessee Eastman process.

captures 85% variation of the internal state of the process. The Hotelling’s T2 statistic was equal to 21.5 at 95% of confidence level. Figure 12 shows that KPCA is able to detect this fault due to exceeding of T2 values beyond the threshold. After a delay of 20 s, there is a sharp jump in Hotelling’s T2 values for this fault 2065

DOI: 10.1021/acs.iecr.6b01916 Ind. Eng. Chem. Res. 2017, 56, 2054−2070

Article

Industrial & Engineering Chemistry Research

Figure 13. (a) BN for Tennessee Eastman and (b) fault propagation pathway for the fault IDV(6).

abnormal condition of FCC and Tennessee Eastman Chemical process. In both case studies, the proposed methodology demonstrated a very powerful diagnostic capability. The strength of the proposed method is it diagnoses root cause of fault as well as shows the propagation pathway of the fault. This information will help operators to take corrective action and recover the process quickly. However, there is still more need for research in this area. For example, most of processes are working in dynamic condition and process variables change with time. In such problems conditional probability of a variable at time t+1 depends on its status at time t and dynamic BN should be implemented for such a problem to handle the dynamic relation of the variables. Also, chemical processes are prone to unknown disturbances. It is not clear how the unobserved disturbances will impact the method in real industrial scenarios. The method needs more tests with industrial data sets.

cooling water inlet temperature will deviate away the temperature of the downstream unit (separator temperature or XMEAS(11)). Because there is no measurement in condenser cooling water inlet temperature, this method is able to find the variable that is mostly affected by the main root (XMEAS(11)) and is able to propagate the fault to other variables. As can be seen in Figure 14b, this deviation will propagate through the network and will influence separator pressure, XMEAS(13), separator level, XMEAS(12), and eventually will effect separator underflow, XMEAS(14).

4. CONCLUSIONS This paper integrates diagnostic information from various diagnostic tools (i.e., KPCA, sensor validation module), combines them with process knowledge using BN, and generates a comprehensive methodology for process FDD. We focused on the different challenges that have received less research focus such as network construction, conditional probability estimation, and loop handling. Sensor faults were separated from process fault using a sensor check module. BN was used to diagnose internal state faults and disturbance faults. The proposed methodology was applied to test different



APPENDIX

Kernel PCA Calculations25,26

Given a sample containing n variables and m measurements, consider the nonlinear training set χ = [x1, x2,... xm]T. An 2066

DOI: 10.1021/acs.iecr.6b01916 Ind. Eng. Chem. Res. 2017, 56, 2054−2070

Article

Industrial & Engineering Chemistry Research

Figure 14. (a) BN for Tennessee Eastman and (b) fault propagation pathway for the fault IDV(12).

χχ T χν = λχν

important property of the feature space is that the dot product of two vectors ⌀i and ⌀j can be calculated as a function of the corresponding vectors xi and xj, that is, T

⌀i ⌀j = k(xi , xj)

Defining ⎡ ⌀ T ⌀ ... ⌀ T ⌀ ⎤ 1 m⎥ ⎢ 1 1 T ⋮ ⎥ K = χχ = ⎢ ⋮ ⋱ ⎢ ⎥ T ⎢⎣ ⌀m ⌀1 ... ⌀mT ⌀m ⎥⎦

(A1)

Assume that the vectors in the feature space are scaled to zero mean and form the training data as χ = [⌀1, ⌀2,... ⌀m]T. Let the sample covariance matrix of the data set in the feature space be S. We have,

⎡ k(x1 , x1) ... k(x1 , xm) ⎤ ⎢ ⎥ ⋱ ⋮ =⎢ ⋮ ⎥ ⎢ ⎥ ⎣ k(xm , x1) ... k(xm , xm)⎦

m

(m − 1)S = χ T χ =

T ∑ ⌀⌀ i i i=1

(A2)

Thus, KPCA in the feature space is equivalent to solving the following eigenvector equation: T ∑ ⌀⌀ i i ν = λν i=1

(A5)

and denoting

α = χν

m

χ T χν =

(A4)

(A6)

we have

(A3)

Kernel trick premultiplies eq A3 by χ:

Kα = λα 2067

(A7) DOI: 10.1021/acs.iecr.6b01916 Ind. Eng. Chem. Res. 2017, 56, 2054−2070

Article

Industrial & Engineering Chemistry Research Equation A7 shows that α and λ are an eigenvector and eigenvalue of K, respectively. To solve ν from eq A6, we premultiply it by χT and use eq A3: χ T α = χ T χν = λν

The T2 Index is as follows:

Index = k̅(zi)T Dk̅(zi)

k̅(zi) is scaled kernel vector. The derivative of the Index with respect to f i is

(A8)

which shows that ν is given by ν = λ−1χ T α

∂k̅(zi) ∂(Index) = 2k̅ T(zi)D ∂fi ∂(fi )

(A9)

Therefore, to calculate the PCA model, we first perform eigendecomposition of eq A7 to obtain λi and αi. Then we use eq A8 to find νi. Considering l principal components, the scores are calculated as t = Λ−1/2PT k(x)

∂ k ̅ (zi ) ∂k(zi) =F ∂fi ∂

∂k(zi) ∂(Index) = 2FDk̅ T(zi) ∂(fi ) ∂fi

k(x) = χϕ = [ ϕ1 ϕ2 ... ϕm ]T ϕ

∂k(zi) ∂k(zi) ∂zi = ∂fi ∂zi ∂fi

(A11)

The T2 is calculated using kernel function as

(zi − xj)T ∂ k(zi , xj) = −2k(zi , xj) c ∂zi

(A13)

Scaling

∂zi = −ξi ∂zi

(A14)

⎡ k(z , x )(z − x )T ⎤ 1 ⎥ ⎢ i 1 i ⎢ T⎥ ∂k(zi) 2 k(z , x )(z − x 2) ⎥ ξ = ⎢ i 1 i ⎥ i ∂fi c⎢ ⋮ ⎢ ⎥ ⎢⎣ k(zi , xm)(zi − xm)T ⎥⎦

Reconstruction-Based Contribution of Variable

The procedure to estimate fault free data by applying a correction in the faulty data is referred to as reconstruction. Reconstruction of the fault free data from faulty measurements can be done by estimating the fault magnitude along the fault direction. The task of fault reconstruction is to estimate the normal values zi, by eliminating the effect of a fault f i from faulty data xi,

⎡ k(z , x )(x − x )T − k(z , x )f ξ T ⎤ i 1 1 i 1 i i ⎥ ⎢ ⎢ T T ⎥ 2 k(zi , x1)(x − x 2) − k(zi , x 2)fi ξi ⎥ = ⎢ ⎥ c⎢ ⋮ ⎢ ⎥ ⎢ k(z , x )(x − x )T − k(z , x )f ξ T ⎥ m i m i i ⎦ ⎣ i m

(A15)

where ξi is the fault direction. The objective of reconstructed based contribution is to find the magnitude f i of a vector with direction such ξi that the fault detection Index of the reconstructed measurement is minimized; that is, we want to find f i such that fi = arg min Index (x − ξifi )

=

2 [Bξi − k(zi)fi ] c

(A25)

Where B is calculated as ⎡ k(z , x )(x − x )T ⎤ 1 ⎥ ⎢ i 1 ⎢ k(z , x )(x − x )T ⎥ 2 ⎥ B=⎢ i 1 ⎥ ⎢ ⋮ ⎥ ⎢ ⎢⎣ k(zi , xm)(x − xm)T ⎥⎦

(A16)

The same concept can be applied to KPCA and find f i such that fi = arg min Index (k(x − ξifi ))

(A24)

Therefore, the vector with the derivative of k(zi) with respect to f i

Where F = I − E, I is the identity matrix, E is an m × m matrix with elements 1/m and 1m is a m dimensional vector whose elements are 1/m.

zi = xi − ξifi

(A23)

and

The calculation of covariance matrix holds if the mapping function in the feature space has zero mean. If this not the case, the vectors in the feature space have to be scaled to zero mean using the sample mean of the training data. The scaling of the kernel vector k(x) is k ̅(x) = [ ϕ1̅ ϕ2̅ ... ϕm̅ ] ϕ ̅ = F[k(x) − K1m ]

(A22)

Because k(zi) = [k(zi, x1), k(zi, x2)...k(zi, xm)] and k(zi, xj) = exp(−(zi − xj)T (zi − xj)/c), we have

(A12)

where

T

(A21)

To calculate the derivative of k(zi) with respect to f i, we know that

= [ ϕ1T ϕ ϕ2T ϕ ... ϕmT ϕ ]

D = PΛ−2PT

(A20)

So, the derivative of Index will be

where P = [αo1... αol ] and Λ = diag{λ1... λl} are the l principal eigenvector and eigenvalues of K and αi = λi αiο and k(x) is

T 2 = k(x)T PΛ−2PT k(x) = k(x)T Dk(x)

(A19)

The scaled kernel vector is k̅(zi) = F[k(zi) − K1m]. Then, we have

(A10)

= [ k(x1 , x) k(x 2 , x) ... k(xm , x)]T

(A18)

(A17) 2068

(A26) DOI: 10.1021/acs.iecr.6b01916 Ind. Eng. Chem. Res. 2017, 56, 2054−2070

Article

Industrial & Engineering Chemistry Research

(15) Zhao, J.; Chen, B.; Shen, J. A Hybrid ANN-ES System for Dynamic Fault Diagnosis of Hydrocracking Process. Comput. Chem. Eng. 1997, 21, 929. (16) Mylaraswamy, D.; Venkatasubramanian, V. A hybrid framework of large scale process fault fiagnosis. Comput. Chem. Eng. 1997, 21, 935. (17) Frank, P. M. Fault Diagnosis in Dynamic Systems Using Analytical and Knowledge-based Redundancy-A Survey and Some New Results. Automatica 1990, 26, 459. (18) Becraft, W. R.; Lee, P. L.; Newell, R. B. Integration of neural networks and expert systems for process fault diagnosis. Proceedings of the 12th International Joint Conference on Artificial Intelligence; Morgan Kaufmann Publishers Inc, 1991. (19) Yu, J.; Rashid, M. M. A Novel Dynamic Bayesian NetworkBased Networked Process Monitoring Approach for Fault Detection, Propagation Identification, and Root Cause Diagnosis. AIChE J. 2013, 59, 2348. (20) Dey, J. S.; Stori, J. A. A Bayesian network approach to root cause diagnosis of process variations. Int. J. Mach. Tool. Manu. 2005, 45, 75. (21) Yu, H.; Khan, F.; Garaniya, V. Modified Independent Component Analysis and Bayesian Network-Based Two-Stage Fault Diagnosis of Process Operations. Ind. Eng. Chem. Res. 2015, 54, 2724. (22) Kobayashi, T.; Simon, D. L. Evaluation of an Enhanced Bank of Kalman Filters for In-Flight Aircraft Engine Sensor Fault Diagnostics. Proceeding of ASME Turbo Expo 2004 Power for Land, Sea, and Air, 2004. (23) Ham, J. H.; Lee, D. D.; Mika, S.; Scholkopf, B. A kernel view of the dimensionality reduction of manifolds. In Proceedings of the TwentyFirst International Conference on Machine Learning, 2004. (24) Scholkopf, B.; Smola, A. Learning with Kernels: Support Vector Machines, Regularization, Optimization, and Beyond; The MIT Press: Cambridge, 2002. (25) Alcala, C. F.; Qin, S. J. Reconstruction-Based Contribution for Process Monitoring with Kernel Principal Component Analysis. Ind. Eng. Chem. Res. 2010, 49, 7849. (26) Qin, S. J. Survey on data-driven industrial process monitoring and diagnosis. Annu. Rev. Control. 2012, 36, 220. (27) Sharma, A. B.; Golubchik, L.; Govindan, R. Sensor Faults: Detection Methods and Prevalencein Real-World Datasets. ACM T. sensor network. 2010, 6, 1. (28) Granger, C. W. J. Investigating Causal Relations by Econometric Models and Cross-spectral Methods. Econometrica 1969, 37, 424. (29) Wiener, N. The Theory of Prediction. In: Modern Mathmatics for Engineers; Beckenbach, E. F., Ed.; McGraw Hill: New York, 1956. (30) Akaike, H. A new look at the statistical model identification. IEEE Trans. Autom. Control 1974, 19, 716. (31) Schwarz, G. Estimating a dimension of a model. Ann. Stat. 1978, 6, 461. (32) Liu, L.; Wan, J. The relationships between Shanghai stock market and CNY/USD exchange rate: New evidence based on crosscorrelation analysis, structural cointegration and nonlinear causality test. Phys. A 2012, 391, 6051. (33) Schreiber, T. Measuring information transfer. Phys. Rev. Lett. 2000, 85, 461. (34) Duan, P.; Shah, S. L.; Chen, T.; Yang, F. A New Information Theory-Based Method for Causality Analysis. ADCONP, Hiroshima, Japan, 2014. (35) Yang, F.; Shah, S. L.; Xiao, D. Signed Directed Graph Modeling of Industrial Processes and their Validation by Data-Based Methods. Conference on Control and Fault Tolerant Systems, Nice, France, 2010. (36) Duan, P.; Yang, F.; Chen, T.; Shah, S. L. Direct Causality Detection via the Transfer Entropy Approach. IEEE Trans. Control Syst. Technol. 2013, 21, 2052. (37) Silverman, B. Density Estimation for Statistics and Data Analysis; Chapman and Hall: London, 1986. (38) Yang, F.; Duan, P.; Shah, S. L.; Chen, T. Capturing Connectivity and Causality in Complex Industrial Processes; Springer Science & Business Media: New York, 2014.

We can now calculate the derivative of index as ∂(Index) 4 = FDk̅ T(zi)[Bξi − k(zi)fi ] ∂fi c

(A27)

After setting the derivative equal to zero and solving for f i, we obtain fi =



ξiT BT FDk̅(zi) k̅ T(zi)FDk̅(zi)

(A28)

AUTHOR INFORMATION

Corresponding Author

*S. A. Imtiaz. E-mail: [email protected]. ORCID

S. A. Imtiaz: 0000-0002-2715-9084 F. Khan: 0000-0002-5638-4299 Notes

The authors declare no competing financial interest.



ACKNOWLEDGMENTS This work was supported by funding from Natural Sciences and Engineering Research Council (NSERC) of Canada under the NSERC Discovery Grant program.



REFERENCES

(1) Dong, J.; Zhang, K.; Huang, Y.; Li, G.; Peng, K. Adaptive total PLS based quality-relevant process monitoring with application to the Tennessee Eastman process. Neurocomputing. 2015, 154, 77. (2) Zhou, Y. Data driven process monitoring based on neural network and classification trees. Ph.D. thesis, Texas A&M University, College Station, 2004. (3) Murray, R. M.; Astrom, K. J.; Boyd, S. P.; Brockett, R. W.; Stein, G. Future Directions in Control in an Information-Rich World. IEEE Conrol Syst. Mag. 2003, 23, 20. (4) Chiang, L. H.; Russell, E. L.; Braatz, R. D. Fault Detection and Diagnosis in Industrial Systems; Springer-Verlag: London, 2001. (5) Venkatasubramanian, V.; Rengaswamy, R.; Kavuri, S. N.; Yin, K. A review of process fault detection and diagnosis: Part iii: Process history based methods. Comput. Chem. Eng. 2003, 27, 327. (6) Venkatasubramanian, V.; Rengaswamy, R.; Yin, K.; Kavuri, S. N. A review of process fault detection and diagnosis Part I: Quantitative model-based methods. Comput. Chem. Eng. 2003, 27, 293. (7) Venkatasubramanian, V.; Rengaswamy, R.; Kavuri, S. N. A review of process fault detection and diagnosis, Part II: Qualitative models and search strategies. Comput. Chem. Eng. 2003, 27, 313. (8) Willsky, A. S.; Jones, H. L. A generalized likelihood ratio approach to state estimation in linear systems subject to abrupt changes. Proc. IEEE Conference on Decision and Control, Phoenix, 1974. (9) Willsky, A. S. A survey of design methods for failure detection systems. Automatica. 1976, 12, 601. (10) Kresta, J. V.; Macgregor, J. F.; Marlin, T. E. Multivariate statistical monitoring of process operation performance. Can. J. Chem. Eng. 1991, 69, 35. (11) Gertler, J. Fault Detection and Diagnoses in Engineering Systems; Marcel Dekker: New York, 1998. (12) Li, W.; Shah, S. Structured residual vector-based approach to sensor fault detection and isolation. J. Process Control 2002, 12, 429. (13) Zhang, Y.; Fan, Y.; Du, W. Nonlinear Process Monitoring Using Regression and Reconstruction Method. IEEE trans. Autom. Sci. Eng. 2016, 13, 1343. (14) Psichogios, D. C.; Ungar, L. H. A hybrid neural network-first principles approach to process modelling. AIChE J. 1992, 38, 1499. 2069

DOI: 10.1021/acs.iecr.6b01916 Ind. Eng. Chem. Res. 2017, 56, 2054−2070

Article

Industrial & Engineering Chemistry Research (39) Mallick, M. R.; Imtiaz, S. A. A Hybrid Method for Process Fault Detection and Diagnosis; The International Federation of Automatic Control: Mumbai, December 2013. (40) McFarlane, R. C.; Reineman, R. C.; Bartee, J. F.; Georgakis, C. Dynamic Simulator for a Model IV Fluid Catalytic Cracking Unit. Comput. Chem. Eng. 1993, 17, 275.

2070

DOI: 10.1021/acs.iecr.6b01916 Ind. Eng. Chem. Res. 2017, 56, 2054−2070